Synopsis

Magnetic Resonance Elastography stiffness estimates in intracranial tumors correlate with intraoperative assessment of tumor consistency, but the spatial kernel-based stiffness calculation of Direct Inversion (DI) creates challenges for small or heterogeneous tumors. The objective of this study is to evaluate an artificial neural network based inversion technique (NNI) in the assessment of small stiff inclusions in a brain phantom. This study shows that NNI can resolve inclusions as small as 1.75cm in diameter with a contrast to noise ratio higher than that of DI. Furthermore, preliminary clinical results show agreement with intraoperative findings.

Introduction

Methods

Image acquisition. MRE was performed on a polyvinyl chloride (PVC) brain phantom with background stiffness of ~3.5kPa (at 60Hz) embedded with 6 spherical inclusions (~7.0kPa at 60Hz) ranging in size from 1.75-3cm at driver power levels of 3, 5, and 7 percent. Two scans were completed at each power level. Scanning was performed on a compact 3T system with a high performance gradient capable of 80 mT/m amplitude and 700 T/m/s^5-7 slew rate using a modified 3D GRE pulse sequence with parameters as follows: 7.2 mT/m motion encoding gradients (MEG), mechanical vibration=60Hz, TE=20.4ms, TR=24.1ms, first-moment nulled MEG, motion encoding sensitivity (MENC) = 7.1 μm/rad, BW = ±25kHz, FOV = 240x240x192mm³, acquisition matrix= 120x120x96, resolution = 2x2x2mm³, flip angle = 12°, 2D ARC reduction=2, 8 channel receiver array (Invivo, Gainesville Fl), scan time = 12:25. One patient with a meningioma (recruited under an IRB-approved protocol) was imaged using the same scan parameters.

Data analysis. After computing the curl of the displacement field, smoothing was performed with a 3x3x3 quartic filter⁸ and stiffness was estimated by algebraic direct inversion (DI)⁹. A 3x3x3 median filter was applied to these stiffness maps to remove outliers. To train the neural network, 1.2 million 7x7x7 simulated data sets were generated with 2-mm isotropic voxels and 4 phase offsets evenly distributed over one period of 60Hz motion. The stiffness of each cube varied randomly between 0.1 and 10 kPa. Sinusoidal waves of the appropriate wavelength were computed from up to 10 point-sources outside of the cube. Noise was added such that the signal to noise ratio varied from 1 to 20 for each example. Model features were the real and imaginary parts of the first harmonic of the curled data selected from a structural element with the same spatial footprint as DI prior to median filtering. An artificial neural network consisting of 3 hidden layers with 24 nodes in each and hyperbolic tangent transfer functions was trained using scaled conjugate gradient backpropagation¹⁰ on 1 million of the simulated datasets, the rest being used for validation. To prevent overfitting, training was stopped when mean-squared error failed to improve in six consecutive iterations in the validation set. When applying NNI to image data, the same features as above for each curl component were fed to NNI, and the final stiffness map (elastogram) was generated as the weighted sum of squares of these images. Contrast to noise ratio was defined as¹¹:

$$CNR=\sqrt{\frac{2 (median_{inclusion} - median_{background})^2}{({σ_{inclusion}}^2+{σ_{background}}^2 )}}$$

For NNI the statistics were calculated directly from the elastogram. For DI, the median and standard deviation were calculated from the filtered map.

Results

Conclusions

Acknowledgements

References